Human detection using RGBD images and Parallel Regional Convolutional networks

Download Article
Ankit Narendrakumar Soni

Abstract- Individuals detection is one of the most sizzling examination points in the domain of PC vision. Albeit deep learning methods, for example, Faster RCNN and SSD have achieved extraordinary accomplishment in individuals detection task utilizing RGB pictures in later a long time, they despise everything experiences the ill effects of the perplexing condition with low enlightenment, complicated background, impediment, and so on. With the development of compact and great depth sensors, for example, Kinect and Xtion, depth data which are invariant to enlightenment and influential to impediment can be acquired effectively now. In this paper, a deep learning method named Parallel RCNN is proposed for individuals detection task utilizing RGB-D data based on Faster RCNN structure. The principle idea of Parallel RCNN is that it takes crude shading picture and encoded depth picture as the contributions of an end-to-end deep neural network at the same time and then concentrates their deep highlights in equal. Through L2 standardization, the highlight maps extracted from every modal data are combined and employed in the following individual’s detection task under the structure of Faster RCNN. To assess the execution of Parallel RCNN, a dataset thoroughly including 2647 RGB-D pictures and 5372 people with hand-labelled comments are created utilizing Kinect v2 sensors. Trial results based on this dataset show that the proposed method can accomplish mean average precision(mAP) of 91.5%, which is 1.5% higher than that of Faster RCNN utilizing RGB pictures.

people detection; Parallel RCNN; deep learning